Computing irregularity for features in medical images
Regular and symmetric forms in nature are associated with health and perfection. In
medical diagnosis the presence of irregular features often suggests an abnormality. Earlier
research identified the irregularity of a lesion border as the most significant factor in
the diagnosis of malignant melanoma. Despite its importance, the visual assessment of
irregularity is difficult. The work described in the thesis was undertaken with the aim
of finding objective computable measures of contour irregularity and their application to
characterisation of pigmented skin lesions.
A descriptive definition of irregularity was formulated, drawing on the related concepts
from the fields of geometry, information theory, statistics, probability, frequency
analysis and pattern theory. It defines irregularity in terms of the five attributes: departure
from a typical sequence (deviation), lack of obvious description, lack of compressibility,
lack of symmetry and lack of a rule for generating a sequence.
Skin lesions were found to be characterised by irregularity due to deviation, lack of
obvious description and lack of a rule. The ellipse was identified as the underlying regular
model and the variability of the lesion boundary was found to be best represented by
Weibull distribution. Three methods for computing irregularity according to the above attributes
were implemented, evaluated and tested. They were based on the Hidden Markov
Models, the Conditional Entropy and the Pattern Theory. The latter method is entirely
novel. Within each method the irregularity measures invariant to the form of boundary
representation and to the parameterisation of each method were identified through experiments
on simulated irregular patterns.
The predictive power of the methods as classifiers of lesion abnormality was tested
on a set of 98 real lesion outlines with histologically confirmed diagnosis. All the methods
showed sensitivity and specificity of over 0.7. The Weibull based Hidden Markov
Models method performed best, with both sensitivity and specificity of 0.82. Ranking
experiments carried out to investigate whether the computed measures corresponds to the
human perception of the border irregularity showed high consistency among the observers
(rank correlation coefficient W=0.886). The correlation between the observations and the
computed measured varied from W=0.49 for the Hidden Markov Models based measure
to W=0.95 for the measure based on the Pattern Theory.